How Long Will the Semiconductor Price Hike Last? Fact-Based Outlook

Prepared: 2026-10-12 (Asia/Tokyo) Scope: The current price surge is led by memory (DRAM, HBM, NAND). A second, smaller hike is coming in foundry wafers (TSMC, from January 2027). "Price hike" below means memory unless stated. Method: Facts come from dated sources, with TrendForce press releases and company earnings commentary preferred. Estimates and probabilities in Sections 4–5 are my own reasoning from those facts, not quotes from any source. Where a date is approximate (marked ~), it is derived from the "age" shown on the search result.

1. Bottom line

2. Verified facts

2.1 Where prices are now

• TrendForce expects conventional DRAM contract prices to rise 10–15% QoQ in Q4 2026 and NAND Flash contract prices to rise 15–20%. It says DRAM stays undersupplied, but the pace of increase is moderating. (TrendForce, 2026-09-30)
• Including HBM, TrendForce puts the broader DRAM rise at 15–20% QoQ in Q4 2026. (Guru3D, citing TrendForce, 2026-09-30)
• Enterprise SSD is the only memory category where price growth is expected to accelerate in Q4, because cloud providers are expanding AI inference. (TrendForce, 2026-09-30)
• The surge began very sharply. Conventional DRAM contract prices rose roughly 93–98% QoQ in Q1 2026, then about 58–63% in Q2. (BuySellRam, citing TrendForce, 2026-10-01)
• TrendForce forecast 13–18% for Q3 2026 and said PC and phone makers were nearing their affordability limit. (BuySellRam, citing TrendForce, 2026-10-01)
• TrendForce has revised forecasts upward more than once. On 2026-06-30 it raised its Q3 PC DRAM outlook from 8–13% to 15–20% and Q4 from 0–5% to 3–8%. The Q4 PC DRAM outlook has since been raised again to 13–18%. (The Build Out, 2026-06-30; PC Master Insider, 2026-10-01)

2.2 Why supply cannot catch up quickly

• HBM (the stacked memory used with AI accelerators) takes roughly three times the wafer area of DDR5 for the same capacity, so AI demand drains capacity from ordinary memory. (BuySellRam, citing Micron, 2026-10-01)
• A modern memory fab typically takes three to five years from groundbreaking to volume output. A finished cleanroom is still about a year from shipping chips. (BuySellRam, 2026-10-01)
• Announced new capacity, with next milestones: (BuySellRam, 2026-10-01)

• SK hynix's M15X fab is expected at full operation around mid-2027, and Samsung's P5 around 2028. (Data Center Dynamics, ~2026-06)
• Samsung's memory chief said on 2026-04-30 that significant shortages should continue through at least 2027, and that some customers already hold allocations through 2027. (Tom's Hardware, 2026-04-30)
• China's CXMT is adding capacity, with Taiwan's Commercial Times expecting about 420,000 wafers a month by 2027, up from about 320,000 now. The same report says China's own AI build-out may absorb much of it. (BuySellRam, citing TrendForce, 2026-10-01)
• TrendForce says memory supply stays tight through 2027 and that the HBM market remains supply-constrained in 2027. It raised its 2027 blended HBM price forecast to +121% YoY. (TrendForce, 2026-09-29)
• PC DRAM supply could fall in 2027 as more capacity shifts to servers. (TrendForce, 2026-09-30)

2.3 What the memory makers say about duration

• Micron's CEO said conditions in fiscal 2027 and 2028 will be much tighter than in 2026. Micron's CFO guided to a "more moderate rate of price increases" for the rest of fiscal 2027. (Yahoo Finance / Stocktwits, 2026-10-01)
• Reported industry sources say Samsung, SK hynix and Micron have sold out their 2027 DRAM and HBM capacity. This is a single-source industry report. (TweakTown, citing DigiTimes, ~2026-08)
• Micron's long-term contracts matter for how prices unwind. It has signed 26 multi-year agreements covering more than 35% of revenue through 2030, and Samsung plans 60–70% of capacity under long-term contracts. (BuySellRam, 2026-10-01)
• Other maker and buyer views, as compiled by BuySellRam from primary links (all as of the dates shown): (BuySellRam, 2026-10-01)

2.4 Independent price-peak forecasts

• Citi expects DRAM and NAND prices to peak in Q2 2027 and then edge down. (BuySellRam, citing Citi, 2026-10-01)
• Gartner said in March that prices might stabilize by Q1 2027. That view now looks early given later upward revisions by TrendForce and Micron. (BuySellRam, citing Gartner, 2026-10-01)
• TrendForce expects server DRAM contract prices to keep rising each quarter through 2H 2027, by smaller amounts. (BuySellRam, citing TrendForce, 2026-10-01)
• Morningstar expects the pricing peak around early 2028, then a steep decline across 2029–31 as supply potentially doubles by 2028. (Yahoo Finance / Stocktwits, 2026-10-01)
• Wolfe Research's Chris Caso said an oversupply is hard to see before 2028 at the very earliest, because the physical cleanroom space does not exist yet. (24/7 Wall St, 2026-07-30)
• A Sungkyunkwan University professor warned that commodity DRAM could reach an oversupply turning point around 2029 if AI returns disappoint. (BigGo Finance, citing the Financial Times, ~2026-07)

2.5 Early signs of strain

• Spot prices for older DDR4 chips and China's used-module market have softened, but contract prices are still rising. (BuySellRam, 2026-10-01)
• IDC expects 2026 smartphone shipments to fall 16.7%, its steepest drop on record, and notebook CPU, DRAM and SSD now account for 68% of parts cost. (BuySellRam, citing IDC, 2026-10-01)
• Hyperscaler spending is still rising, but a growing share is debt-funded. Goldman counted $489 billion of AI-related bond and loan issuance by late July. (BuySellRam, citing Goldman/Axios, 2026-10-01)
• TrendForce estimates DRAM and NAND are 47% of major cloud providers' capex in 2026 and 68% in 2027, so part of record AI spending is just record memory prices. (BuySellRam, citing TrendForce, 2026-10-01)

2.6 Foundry: a separate, smaller hike

• Nikkei Asia reported that TSMC plans base wafer price increases of 5–10% from the start of 2027, covering advanced and mature nodes. HPC orders above a customer's original forecast may carry a further 10–15% premium. (ChainCatcher, citing Nikkei, 2026-07-21)
• Mature nodes (12nm, 16nm, 28nm) reportedly face up to 10%, the first broad mature-node increase in over three years. This is single-sourced via unnamed sources and TSMC has not confirmed it. (AI Weekly, ~2026-07)
• TSMC's chairman said the company avoids sudden price jumps. (Tom's Hardware, ~2026-07)

3. Reasoning

1. The rate of increase is falling, but the level is still climbing. Conventional DRAM contract growth went from roughly 95% (Q1) to about 60% (Q2) to a forecast 13–18% (Q3) to a forecast 10–15% (Q4). Chaining the midpoints, and using the Q3 forecast because I did not verify the actual, implies conventional DRAM contract prices of roughly 4x their Q4 2025 level by end of 2026. This is my arithmetic, not a sourced figure.
2. Price peaks when QoQ growth reaches zero, which is later than "slowing." Every source that expects a turn expects it after growth fades, not now. Citi's Q2 2027 peak is the earliest credible date among the named forecasts.
3. Supply response is mechanically slow. The first meaningful new output (M15X, Micron ID1, Yongin cleanroom) lands in mid-2027, and each needs a ramp period. Most volume arrives from late 2027 to 2029. This sets a floor: broad price relief before 2H 2027 is unlikely.
4. Contracts lock in tightness. Large 2027 volumes are already committed, and multi-year deals reportedly include price floors. That slows both the rise and the eventual fall, and pushes the unwind onto spot buyers first.
5. Forecasts have repeatedly been revised up. TrendForce raised Q3 and Q4 outlooks in June and again in September, and Micron and Samsung have each pushed their end dates later. A smooth deceleration path is therefore the optimistic case, and upside surprises have been the pattern.
6. Watch the incentives. Makers gain contract leverage from a long-shortage narrative, and PC makers gain from a short one (BuySellRam makes the same point). I weight the observable data (contract prices, fab dates, committed volumes) above the headline end-dates.
7. The main risk to a long hike is AI financing, not memory supply. If hyperscaler budgets flatten while new fabs arrive, the 2019 pattern could repeat sooner than 2028.

4. Scenarios (my estimates)

Probabilities are subjective judgments. They are not sourced and carry real uncertainty.

By product, base case:

5. Signals that would change this view

6. Caveats

• Several price-peak forecasts (Gartner, Citi, AMD) are cited second-hand through BuySellRam, which is a reseller with its own incentives. I used its primary links where available, but I did not open every one.
• The Q3 2026 actual contract-price change was not verified. Only forecasts were found.
• The TSMC increase is single-sourced via Nikkei and unconfirmed by TSMC.
• Dates marked ~ are approximate.
• This is analysis for planning, not investment or financial advice.

7. Source list